[clang] [llvm] [hlsl][dxil][spirv] Add `fma` intrinsic (PR #185304)
via cfe-commits
cfe-commits at lists.llvm.org
Sun Mar 8 10:19:58 PDT 2026
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-directx
Author: None (NeKon69)
<details>
<summary>Changes</summary>
This PR adds `fma` HLSL intrinsic (supports matrices in dxil backend)
It follows all of the steps from #<!-- -->99117 and also closes that issue.
---
Patch is 32.05 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/185304.diff
16 Files Affected:
- (modified) clang/include/clang/Basic/Builtins.td (+6)
- (modified) clang/include/clang/Basic/DiagnosticSemaKinds.td (+6)
- (modified) clang/lib/CodeGen/CGHLSLBuiltins.cpp (+16)
- (modified) clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h (+16)
- (modified) clang/lib/Sema/SemaHLSL.cpp (+61)
- (added) clang/test/CodeGen/SPIRV/hlsl-intrinsics/fma.ll ()
- (added) clang/test/CodeGenHLSL/builtins/fma.hlsl (+151)
- (added) clang/test/Sema/incompatible-function-to-ptr-decay.c (+18)
- (added) clang/test/SemaHLSL/BuiltIns/fma-errors.hlsl (+145)
- (modified) llvm/include/llvm/IR/IntrinsicsDirectX.td (+2)
- (modified) llvm/include/llvm/IR/IntrinsicsSPIRV.td (+1-1)
- (modified) llvm/lib/Target/DirectX/DXIL.td (+10)
- (modified) llvm/lib/Target/DirectX/DXILShaderFlags.cpp (+10-1)
- (modified) llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp (+2)
- (modified) llvm/test/CodeGen/DirectX/ShaderFlags/double-extensions.ll (+8)
- (added) llvm/test/CodeGen/SPIRV/hlsl-intrinsics/fma.ll (+53)
``````````diff
diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td
index 531c3702161f2..542249f829424 100644
--- a/clang/include/clang/Basic/Builtins.td
+++ b/clang/include/clang/Basic/Builtins.td
@@ -5318,6 +5318,12 @@ def HLSLNormalize : LangBuiltin<"HLSL_LANG"> {
let Prototype = "void(...)";
}
+def HLSLFma : LangBuiltin<"HLSL_LANG"> {
+ let Spellings = ["__builtin_hlsl_elementwise_fma"];
+ let Attributes = [NoThrow, Const, CustomTypeChecking];
+ let Prototype = "void(...)";
+}
+
def HLSLRcp : LangBuiltin<"HLSL_LANG"> {
let Spellings = ["__builtin_hlsl_elementwise_rcp"];
let Attributes = [NoThrow, Const, CustomTypeChecking];
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index 8882ac9b8c0a8..787cd7bcc61bb 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -13229,6 +13229,12 @@ def err_builtin_invalid_arg_type: Error<
"%plural{0:|: }3"
"%plural{[0,3]:type|:types}1 (was %4)">;
+def err_builtin_requires_double_type: Error<
+ "%ordinal0 argument must be a scalar, vector, or matrix of double type (was %1)">;
+
+def err_builtin_requires_fp_scalar_or_vector_type: Error<
+ "%ordinal0 argument must be a scalar or vector of floating-point type (was %1)">;
+
def err_bswapg_invalid_bit_width : Error<
"_BitInt type %0 (%1 bits) must be a multiple of 16 bits for byte swapping">;
diff --git a/clang/lib/CodeGen/CGHLSLBuiltins.cpp b/clang/lib/CodeGen/CGHLSLBuiltins.cpp
index 70891eac39425..bb5eaf12c93cc 100644
--- a/clang/lib/CodeGen/CGHLSLBuiltins.cpp
+++ b/clang/lib/CodeGen/CGHLSLBuiltins.cpp
@@ -979,6 +979,22 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
retType, CGM.getHLSLRuntime().getIsNaNIntrinsic(),
ArrayRef<Value *>{Op0}, nullptr, "hlsl.isnan");
}
+ case Builtin::BI__builtin_hlsl_elementwise_fma: {
+ Value *M = EmitScalarExpr(E->getArg(0));
+ Value *A = EmitScalarExpr(E->getArg(1));
+ Value *B = EmitScalarExpr(E->getArg(2));
+ if (CGM.getTarget().getTriple().isDXIL())
+ return Builder.CreateIntrinsic(M->getType(), Intrinsic::dx_fma,
+ ArrayRef<Value *>{M, A, B}, nullptr,
+ "dx.fma");
+
+ if (CGM.getTarget().getTriple().isSPIRV())
+ return Builder.CreateIntrinsic(M->getType(), Intrinsic::spv_fma,
+ ArrayRef<Value *>{M, A, B}, nullptr,
+ "spv.fma");
+
+ break;
+ }
case Builtin::BI__builtin_hlsl_mad: {
Value *M = EmitScalarExpr(E->getArg(0));
Value *A = EmitScalarExpr(E->getArg(1));
diff --git a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
index 2543401bdfbf9..ab5c6edd6d555 100644
--- a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h
@@ -1891,6 +1891,22 @@ float3 pow(float3, float3);
_HLSL_BUILTIN_ALIAS(__builtin_elementwise_pow)
float4 pow(float4, float4);
+//===----------------------------------------------------------------------===//
+// fused multiply-add builtins
+//===----------------------------------------------------------------------===//
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_fma)
+double fma(double, double, double);
+
+template <int s>
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_fma)
+vector<double, s> fma(vector<double, s>, vector<double, s>, vector<double, s>);
+
+template <int w, int h>
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_elementwise_fma)
+matrix<double, w, h> fma(matrix<double, w, h>, matrix<double, w, h>,
+ matrix<double, w, h>);
+
//===----------------------------------------------------------------------===//
// reversebits builtins
//===----------------------------------------------------------------------===//
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index 804ea70aaddce..624f621b532a1 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -31,6 +31,7 @@
#include "clang/Basic/TargetInfo.h"
#include "clang/Sema/Initialization.h"
#include "clang/Sema/Lookup.h"
+#include "clang/Sema/Ownership.h"
#include "clang/Sema/ParsedAttr.h"
#include "clang/Sema/Sema.h"
#include "clang/Sema/Template.h"
@@ -3040,6 +3041,36 @@ static bool CheckFloatOrHalfRepresentation(Sema *S, SourceLocation Loc,
return false;
}
+static bool CheckFloatOrHalfOrDoubleRepresentation(Sema *S, SourceLocation Loc,
+ int ArgOrdinal,
+ clang::QualType PassedType) {
+ clang::QualType BaseType =
+ PassedType->isVectorType()
+ ? PassedType->castAs<clang::VectorType>()->getElementType()
+ : PassedType;
+ if (!BaseType->isFloatingType())
+ return S->Diag(Loc, diag::err_builtin_requires_fp_scalar_or_vector_type)
+ << ArgOrdinal << PassedType;
+ return false;
+}
+
+static bool CheckAnyDoubleRepresentation(Sema *S, SourceLocation Loc,
+ int ArgOrdinal,
+ clang::QualType PassedType) {
+ clang::QualType BaseType =
+ PassedType->isVectorType()
+ ? PassedType->castAs<clang::VectorType>()->getElementType()
+ : PassedType->isMatrixType()
+ ? PassedType->castAs<clang::MatrixType>()->getElementType()
+ : PassedType;
+ if (!BaseType->isDoubleType()) {
+ return S->Diag(Loc, diag::err_builtin_requires_double_type)
+ << ArgOrdinal << PassedType;
+ }
+
+ return false;
+}
+
static bool CheckModifiableLValue(Sema *S, CallExpr *TheCall,
unsigned ArgIndex) {
auto *Arg = TheCall->getArg(ArgIndex);
@@ -3787,6 +3818,35 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
TheCall->setType(ArgTyA);
break;
}
+ case Builtin::BI__builtin_hlsl_elementwise_fma: {
+ if (SemaRef.checkArgCount(TheCall, 3)) {
+ return true;
+ }
+ const llvm::Triple &TT = getASTContext().getTargetInfo().getTriple();
+ // This check is here because emitting a general error for both backends
+ // here (like for exmaple "Accepts only floating points") won't end really
+ // good. after that we still need to check if the types satisfy
+ // backends constrains, so we better check everything now rather than
+ // confusing user with 2 different error messages
+
+ if (TT.isSPIRV()) {
+ // SPIR-V accept any float (besides matrices)
+ if (CheckAllArgTypesAreCorrect(&SemaRef, TheCall,
+ CheckFloatOrHalfOrDoubleRepresentation))
+ return true;
+ } else if (TT.isDXIL()) {
+ // while DirectX accepts only double
+ if (CheckAllArgTypesAreCorrect(&SemaRef, TheCall,
+ CheckAnyDoubleRepresentation))
+ return true;
+ }
+
+ ExprResult A = TheCall->getArg(0);
+ QualType ArgTyA = A.get()->getType();
+ // return type is the same as input type
+ TheCall->setType(ArgTyA);
+ break;
+ }
case Builtin::BI__builtin_hlsl_elementwise_sign: {
if (SemaRef.PrepareBuiltinElementwiseMathOneArgCall(TheCall))
return true;
@@ -3936,6 +3996,7 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
case Builtin::BI__builtin_elementwise_exp10:
case Builtin::BI__builtin_elementwise_floor:
case Builtin::BI__builtin_elementwise_fmod:
+ case Builtin::BI__builtin_elementwise_fma:
case Builtin::BI__builtin_elementwise_log:
case Builtin::BI__builtin_elementwise_log2:
case Builtin::BI__builtin_elementwise_log10:
diff --git a/clang/test/CodeGen/SPIRV/hlsl-intrinsics/fma.ll b/clang/test/CodeGen/SPIRV/hlsl-intrinsics/fma.ll
new file mode 100644
index 0000000000000..e69de29bb2d1d
diff --git a/clang/test/CodeGenHLSL/builtins/fma.hlsl b/clang/test/CodeGenHLSL/builtins/fma.hlsl
new file mode 100644
index 0000000000000..88b8e27c37043
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/fma.hlsl
@@ -0,0 +1,151 @@
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
+// RUN: dxil-pc-shadermodel6.3-library %s -DTEST_DXIL \
+// RUN: -fmatrix-memory-layout=row-major -emit-llvm -disable-llvm-passes -o - | \
+// RUN: FileCheck %s --check-prefixes=CHECK,DXIL_CHECK -DTARGET=dx
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
+// RUN: spirv-unknown-vulkan-compute %s -DTEST_SPIRV \
+// RUN: -fmatrix-memory-layout=row-major -emit-llvm -disable-llvm-passes -o - | \
+// RUN: FileCheck %s --check-prefixes=CHECK,SPIRV_CHECK -DTARGET=spv
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
+// RUN: spirv-unknown-vulkan-compute %s -DTEST_SPIRV_HALF -fnative-half-type \
+// RUN: -fmatrix-memory-layout=row-major -emit-llvm -disable-llvm-passes -o - | \
+// RUN: FileCheck %s --check-prefix=SPIRV_HALF_CHECK
+
+// CHECK-LABEL: define {{.*}} double @{{.*}}fma_double{{.*}}(
+// CHECK: %[[P0:.*]] = load double, ptr %{{.*}}, align 8
+// CHECK: %[[P1:.*]] = load double, ptr %{{.*}}, align 8
+// CHECK: %[[P2:.*]] = load double, ptr %{{.*}}, align 8
+// CHECK: %{{dx|spv}}.fma = call reassoc nnan ninf nsz arcp afn double @llvm.[[TARGET]].fma.f64(double %[[P0]], double %[[P1]], double %[[P2]])
+// CHECK: ret double %{{dx|spv}}.fma
+double dxil_fma_double(double a, double b, double c) { return fma(a, b, c); }
+
+// CHECK-LABEL: define {{.*}} <2 x double> @{{.*}}fma_double2{{.*}}(
+// CHECK: %[[P0:.*]] = load <2 x double>, ptr %{{.*}}, align 16
+// CHECK: %[[P1:.*]] = load <2 x double>, ptr %{{.*}}, align 16
+// CHECK: %[[P2:.*]] = load <2 x double>, ptr %{{.*}}, align 16
+// CHECK: %{{dx|spv}}.fma = call reassoc nnan ninf nsz arcp afn <2 x double> @llvm.[[TARGET]].fma.v2f64(<2 x double> %[[P0]], <2 x double> %[[P1]], <2 x double> %[[P2]])
+// CHECK: ret <2 x double> %{{dx|spv}}.fma
+double2 dxil_fma_double2(double2 a, double2 b, double2 c) { return fma(a, b, c); }
+
+// CHECK-LABEL: define {{.*}} <3 x double> @{{.*}}fma_double3{{.*}}(
+// CHECK: %[[P0:.*]] = load <3 x double>, ptr %{{.*}}, align 32
+// CHECK: %[[P1:.*]] = load <3 x double>, ptr %{{.*}}, align 32
+// CHECK: %[[P2:.*]] = load <3 x double>, ptr %{{.*}}, align 32
+// CHECK: %{{dx|spv}}.fma = call reassoc nnan ninf nsz arcp afn <3 x double> @llvm.[[TARGET]].fma.v3f64(<3 x double> %[[P0]], <3 x double> %[[P1]], <3 x double> %[[P2]])
+// CHECK: ret <3 x double> %{{dx|spv}}.fma
+double3 dxil_fma_double3(double3 a, double3 b, double3 c) { return fma(a, b, c); }
+
+// CHECK-LABEL: define {{.*}} <4 x double> @{{.*}}fma_double4{{.*}}(
+// CHECK: %[[P0:.*]] = load <4 x double>, ptr %{{.*}}, align 32
+// CHECK: %[[P1:.*]] = load <4 x double>, ptr %{{.*}}, align 32
+// CHECK: %[[P2:.*]] = load <4 x double>, ptr %{{.*}}, align 32
+// CHECK: %{{dx|spv}}.fma = call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.[[TARGET]].fma.v4f64(<4 x double> %[[P0]], <4 x double> %[[P1]], <4 x double> %[[P2]])
+// CHECK: ret <4 x double> %{{dx|spv}}.fma
+double4 dxil_fma_double4(double4 a, double4 b, double4 c) { return fma(a, b, c); }
+
+#ifdef TEST_DXIL
+
+// DXIL_CHECK-LABEL: define {{.*}} <4 x double> @{{.*}}dxil_fma_double1x4{{.*}}(
+// DXIL_CHECK: %dx.fma = call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.dx.fma.v4f64(
+// DXIL_CHECK: ret <4 x double> %dx.fma
+double1x4 dxil_fma_double1x4(double1x4 a, double1x4 b, double1x4 c) { return fma(a, b, c); }
+
+// DXIL_CHECK-LABEL: define {{.*}} <4 x double> @{{.*}}dxil_fma_double4x1{{.*}}(
+// DXIL_CHECK: %dx.fma = call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.dx.fma.v4f64(
+// DXIL_CHECK: ret <4 x double> %dx.fma
+double4x1 dxil_fma_double4x1(double4x1 a, double4x1 b, double4x1 c) { return fma(a, b, c); }
+
+// DXIL_CHECK-LABEL: define {{.*}} <4 x double> @{{.*}}dxil_fma_double2x2{{.*}}(
+// DXIL_CHECK: %dx.fma = call reassoc nnan ninf nsz arcp afn <4 x double> @llvm.dx.fma.v4f64(
+// DXIL_CHECK: ret <4 x double> %dx.fma
+double2x2 dxil_fma_double2x2(double2x2 a, double2x2 b, double2x2 c) { return fma(a, b, c); }
+
+// DXIL_CHECK-LABEL: define {{.*}} <6 x double> @{{.*}}dxil_fma_double2x3{{.*}}(
+// DXIL_CHECK: %dx.fma = call reassoc nnan ninf nsz arcp afn <6 x double> @llvm.dx.fma.v6f64(
+// DXIL_CHECK: ret <6 x double> %dx.fma
+double2x3 dxil_fma_double2x3(double2x3 a, double2x3 b, double2x3 c) { return fma(a, b, c); }
+
+// DXIL_CHECK-LABEL: define {{.*}} <6 x double> @{{.*}}dxil_fma_double3x2{{.*}}(
+// DXIL_CHECK: %dx.fma = call reassoc nnan ninf nsz arcp afn <6 x double> @llvm.dx.fma.v6f64(
+// DXIL_CHECK: ret <6 x double> %dx.fma
+double3x2 dxil_fma_double3x2(double3x2 a, double3x2 b, double3x2 c) { return fma(a, b, c); }
+
+// DXIL_CHECK-LABEL: define {{.*}} <9 x double> @{{.*}}dxil_fma_double3x3{{.*}}(
+// DXIL_CHECK: %dx.fma = call reassoc nnan ninf nsz arcp afn <9 x double> @llvm.dx.fma.v9f64(
+// DXIL_CHECK: ret <9 x double> %dx.fma
+double3x3 dxil_fma_double3x3(double3x3 a, double3x3 b, double3x3 c) { return fma(a, b, c); }
+
+// DXIL_CHECK-LABEL: define {{.*}} <16 x double> @{{.*}}dxil_fma_double4x4{{.*}}(
+// DXIL_CHECK: %dx.fma = call reassoc nnan ninf nsz arcp afn <16 x double> @llvm.dx.fma.v16f64(
+// DXIL_CHECK: ret <16 x double> %dx.fma
+double4x4 dxil_fma_double4x4(double4x4 a, double4x4 b, double4x4 c) { return fma(a, b, c); }
+#endif
+
+#ifdef TEST_SPIRV
+// SPIRV_CHECK-LABEL: define {{.*}} float @{{.*}}spv_fma_float{{.*}}(
+// SPIRV_CHECK: %[[P0:.*]] = load float, ptr %{{.*}}, align 4
+// SPIRV_CHECK: %[[P1:.*]] = load float, ptr %{{.*}}, align 4
+// SPIRV_CHECK: %[[P2:.*]] = load float, ptr %{{.*}}, align 4
+// SPIRV_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn float @llvm.spv.fma.f32(float %[[P0]], float %[[P1]], float %[[P2]])
+// SPIRV_CHECK: ret float %spv.fma
+float spv_fma_float(float a, float b, float c) { return fma(a, b, c); }
+
+// SPIRV_CHECK-LABEL: define {{.*}} <2 x float> @{{.*}}spv_fma_float2{{.*}}(
+// SPIRV_CHECK: %[[P0:.*]] = load <2 x float>, ptr %{{.*}}, align 8
+// SPIRV_CHECK: %[[P1:.*]] = load <2 x float>, ptr %{{.*}}, align 8
+// SPIRV_CHECK: %[[P2:.*]] = load <2 x float>, ptr %{{.*}}, align 8
+// SPIRV_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn <2 x float> @llvm.spv.fma.v2f32(<2 x float> %[[P0]], <2 x float> %[[P1]], <2 x float> %[[P2]])
+// SPIRV_CHECK: ret <2 x float> %spv.fma
+float2 spv_fma_float2(float2 a, float2 b, float2 c) { return fma(a, b, c); }
+
+// SPIRV_CHECK-LABEL: define {{.*}} <3 x float> @{{.*}}spv_fma_float3{{.*}}(
+// SPIRV_CHECK: %[[P0:.*]] = load <3 x float>, ptr %{{.*}}, align 16
+// SPIRV_CHECK: %[[P1:.*]] = load <3 x float>, ptr %{{.*}}, align 16
+// SPIRV_CHECK: %[[P2:.*]] = load <3 x float>, ptr %{{.*}}, align 16
+// SPIRV_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn <3 x float> @llvm.spv.fma.v3f32(<3 x float> %[[P0]], <3 x float> %[[P1]], <3 x float> %[[P2]])
+// SPIRV_CHECK: ret <3 x float> %spv.fma
+float3 spv_fma_float3(float3 a, float3 b, float3 c) { return fma(a, b, c); }
+
+// SPIRV_CHECK-LABEL: define {{.*}} <4 x float> @{{.*}}spv_fma_float4{{.*}}(
+// SPIRV_CHECK: %[[P0:.*]] = load <4 x float>, ptr %{{.*}}, align 16
+// SPIRV_CHECK: %[[P1:.*]] = load <4 x float>, ptr %{{.*}}, align 16
+// SPIRV_CHECK: %[[P2:.*]] = load <4 x float>, ptr %{{.*}}, align 16
+// SPIRV_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn <4 x float> @llvm.spv.fma.v4f32(<4 x float> %[[P0]], <4 x float> %[[P1]], <4 x float> %[[P2]])
+// SPIRV_CHECK: ret <4 x float> %spv.fma
+float4 spv_fma_float4(float4 a, float4 b, float4 c) { return fma(a, b, c); }
+
+#endif
+
+#ifdef TEST_SPIRV_HALF
+// SPIRV_HALF_CHECK-LABEL: define {{.*}} half @{{.*}}spv_fma_half{{.*}}(
+// SPIRV_HALF_CHECK: %[[P0:.*]] = load half, ptr %{{.*}}, align 2
+// SPIRV_HALF_CHECK: %[[P1:.*]] = load half, ptr %{{.*}}, align 2
+// SPIRV_HALF_CHECK: %[[P2:.*]] = load half, ptr %{{.*}}, align 2
+// SPIRV_HALF_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fma.f16(half %[[P0]], half %[[P1]], half %[[P2]])
+// SPIRV_HALF_CHECK: ret half %spv.fma
+half spv_fma_half(half a, half b, half c) { return fma(a, b, c); }
+
+// SPIRV_HALF_CHECK-LABEL: define {{.*}} <2 x half> @{{.*}}spv_fma_half2{{.*}}(
+// SPIRV_HALF_CHECK: %[[P0:.*]] = load <2 x half>, ptr %{{.*}}, align 4
+// SPIRV_HALF_CHECK: %[[P1:.*]] = load <2 x half>, ptr %{{.*}}, align 4
+// SPIRV_HALF_CHECK: %[[P2:.*]] = load <2 x half>, ptr %{{.*}}, align 4
+// SPIRV_HALF_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn <2 x half> @llvm.spv.fma.v2f16(<2 x half> %[[P0]], <2 x half> %[[P1]], <2 x half> %[[P2]])
+// SPIRV_HALF_CHECK: ret <2 x half> %spv.fma
+half2 spv_fma_half2(half2 a, half2 b, half2 c) { return fma(a, b, c); }
+
+// SPIRV_HALF_CHECK-LABEL: define {{.*}} <3 x half> @{{.*}}spv_fma_half3{{.*}}(
+// SPIRV_HALF_CHECK: %[[P0:.*]] = load <3 x half>, ptr %{{.*}}, align 8
+// SPIRV_HALF_CHECK: %[[P1:.*]] = load <3 x half>, ptr %{{.*}}, align 8
+// SPIRV_HALF_CHECK: %[[P2:.*]] = load <3 x half>, ptr %{{.*}}, align 8
+// SPIRV_HALF_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn <3 x half> @llvm.spv.fma.v3f16(<3 x half> %[[P0]], <3 x half> %[[P1]], <3 x half> %[[P2]])
+// SPIRV_HALF_CHECK: ret <3 x half> %spv.fma
+half3 spv_fma_half3(half3 a, half3 b, half3 c) { return fma(a, b, c); }
+
+// SPIRV_HALF_CHECK-LABEL: define {{.*}} <4 x half> @{{.*}}spv_fma_half4{{.*}}(
+// SPIRV_HALF_CHECK: %[[P0:.*]] = load <4 x half>, ptr %{{.*}}, align 8
+// SPIRV_HALF_CHECK: %[[P1:.*]] = load <4 x half>, ptr %{{.*}}, align 8
+// SPIRV_HALF_CHECK: %[[P2:.*]] = load <4 x half>, ptr %{{.*}}, align 8
+// SPIRV_HALF_CHECK: %spv.fma = call reassoc nnan ninf nsz arcp afn <4 x half> @llvm.spv.fma.v4f16(<4 x half> %[[P0]], <4 x half> %[[P1]], <4 x half> %[[P2]])
+// SPIRV_HALF_CHECK: ret <4 x half> %spv.fma
+half4 spv_fma_half4(half4 a, half4 b, half4 c) { return fma(a, b, c); }
+#endif
diff --git a/clang/test/Sema/incompatible-function-to-ptr-decay.c b/clang/test/Sema/incompatible-function-to-ptr-decay.c
new file mode 100644
index 0000000000000..240b5b8763a23
--- /dev/null
+++ b/clang/test/Sema/incompatible-function-to-ptr-decay.c
@@ -0,0 +1,18 @@
+// RUN: %clang_cc1 -fsyntax-only -fexperimental-overflow-behavior-types -verify %s
+
+// Issue 182534
+int foo();
+
+void bar(__attribute__((opencl_global)) int*); // #cldecl
+void baz(__ob_wrap int*); // #ofdecl
+
+void a() {
+ bar(foo);
+ // expected-error at -1 {{passing 'int (*)()' to parameter of type '__global int *' changes address space of pointer}}
+ // expected-note@#cldecl {{passing argument to parameter here}}
+ __ob_trap int val[10];
+ baz(val);
+ // expected-error at -1 {{assigning to '__ob_wrap int *' from '__ob_trap int *' with incompatible overflow behavior types ('__ob_wrap' and '__ob_trap')}}
+ // expected-note@#ofdecl {{passing argument to parameter here}}
+}
+
diff --git a/clang/test/SemaHLSL/BuiltIns/fma-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/fma-errors.hlsl
new file mode 100644
index 0000000000000..1ed7b34b4396f
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/fma-errors.hlsl
@@ -0,0 +1,145 @@
+// RUN: %clang_cc1 -finclude-default-header -fnative-half-type -x hlsl \
+// RUN: -triple dxil-pc-shadermodel6.6-library %s -DTEST_DXIL \
+// RUN: -emit-llvm-only -disable-llvm-passes -verify=dxil
+// RUN: %clang_cc1 -finclude-default-header -fnative-half-type -x hlsl \
+// RUN: -triple spirv-unknown-vulkan-compute %s -DTEST_SPIRV \
+// RUN: -emit-llvm-only -disable-llvm-passes -verify=spv
+
+#ifdef TEST_DXIL
+float dxil_fma_float(float a, float b, float c) {
+ return fma(a, b, c);
+ // dxil-error at -1 {{1st argument must be a scalar, vector, or matrix of double type (was 'float')}}
+}
+
+float2 dxil_fma_float2(float2 a, float2 b, float2 c) {
+ return fma(a, b, c);
+ // dxil-error at -1 {{1st argument must be a scalar, vector, or matrix of double type (was 'float2' (aka 'vector<float, 2>'))}}
+}
+
+float4 dxil_fma_float4(float4 a, float4 b, float4 c) {
+ return fma(a, b, c);
+ // dxil-error at -1 {{1st argument must be a scalar, vector, or matrix of double type (was 'float4' (aka 'vector<float, 4>'))}}
+}
+
+float2x2 dxil_fma_float2x2(float2x2 a, float2x2 b, float2x2 c) {
+ return fma(a, b, c);
+ // dxil-error at -1 {{1st argument must be a scalar, vector, or matrix of double type (was 'float2x2' (aka 'matrix<float, 2, 2>'))}}
+}
+
+doub...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/185304
More information about the cfe-commits
mailing list